Evaluating the Impact of Low-Light Image Enhancement Methods on

Runner Re-Identiﬁcation in the Wild

Oliverio J. Santana

, Javier Lorenzo-Navarro

, David Freire-Obreg

Daniel Hern

andez-Sosa

and Modesto Castrill

on-Santana

SIANI, Universidad de Las Palmas de Gran Canaria, Spain

Keywords:

Computer Vision, Person Re-Identiﬁcation, in the Wild Sporting Events, Low-Light Image Enhancement.

Abstract:

Person re-identiﬁcation (ReID) is a trending topic in computer vision. Signiﬁcant developments have been

achieved, but most rely on datasets with subjects captured statically within a short period of time in rather good

lighting conditions. In the wild scenarios, such as long-distance races that involve widely varying lighting

conditions, from full daylight to night, present a considerable challenge. This issue cannot be addressed

by increasing the exposure time on the capture device, as the runners’ motion will lead to blurred images,

hampering any ReID attempts. In this paper, we survey some low-light image enhancement methods. Our

results show that including an image processing step in a ReID pipeline before extracting the distinctive body

appearance features from the subjects can provide signiﬁcant performance improvements.

1 INTRODUCTION

Our ability to effortlessly identify all subjects of an

image relies on a solid semantic understanding of

the people in the scene. However, despite the hu-

man capability, this ability remains a challenge for our

state-of-the-art visual recognition systems. One of the

primary goals in this area is person re-identiﬁcation

(ReID), where the task is to match subjects captured

in different spots and times.

ReID research has seen signiﬁcant progress in the

last few years (Penate-Sanchez et al., 2020; Ning

et al., 2021). New challenging in the wild datasets

have emerged due to the evolution of capture de-

vices, going beyond short-time ReID with homoge-

neous illumination conditions. These new collec-

tions have shown several compelling problems in tra-

ditional ReID benchmarks. In this regard, long-term

ReID copes with substantial variability in space and

time. For instance, individuals may be recorded not

uniquely affected by pose and occlusion variations but

also strongly forced due to different resolutions, i.e.,

producing multi-scale detections, appearance incon-

https://orcid.org/0000-0001-7511-5783

https://orcid.org/0000-0002-2834-2067

https://orcid.org/0000-0003-2378-4277

https://orcid.org/0000-0003-3022-7698

https://orcid.org/0000-0002-8673-2725

sistencies due to clothing change, and many environ-

mental and lighting variations.

State-of-the-art face recognition approaches ap-

plied in surveillance and standard ReID scenarios

have evidenced low performance (Cheng et al., 2020;

Dietlmeier et al., 2020) due to the need for high reso-

lution and image quality. Lately, low-light image en-

hancement has attracted the community’s attention on

this subject to cope with illumination changes in the

target images.

Poorly illuminated images suffer from low con-

trast and high level of noise (Rahman et al., 2021).

Signiﬁcant issues like over-saturation are introduced

on image regions with high-intensity pixels by

straightly adjusting the illumination property. Conse-

quently, the community has proposed several image

enhancement methods to address this problem.

Traditional low-light image enhancement meth-

ods rely on the Retinex Theory (Land, 1977), which

focuses on an image’s dynamic range and color con-

stancy and recovers the contrast by accurately esti-

mating illumination.

Nonetheless, some of these methods may gener-

ate color distortion or they are prone to enhance im-

age brightness presenting unnatural effects (He et al.,

2020). More recently, deep learning techniques have

pushed advances on this challenging task (Zhai et al.,

2021) in two directions: image quality and processing

time decrease.

Santana, O., Lorenzo-Navarro, J., Freire-Obregón, D., Hernández-Sosa, D. and Castrillón-Santana, M.

Evaluating the Impact of Low-Light Image Enhancement Methods on Runner Re-Identiﬁcation in the Wild.

DOI: 10.5220/0011652000003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 641-648

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

641

However, despite encouraging progress on low-

light image enhancement, we believe there is room for

improvement. For instance, only a few recent works

have been tested on video clips (Lv et al., 2018; Li

et al., 2022). Unlike still images, video clips present

a continuous lighting signal variation that may affect

subsequent processes, such as person ReID.

This work takes a step towards image enhance-

ment evaluation in dynamic scenarios, proposing a

novel ReID pipeline to unify both tasks, image en-

hancement and ReID. Both are computed differently;

the former exploits the complete scene to enhance

the image by using as much information as possible,

whereas the latter focuses on a speciﬁc region of in-

terest of the improved image.

Additionally, we developed a ReID assessment

considering different state-of-the-art methods for im-

age enhancement. In this regard, traditional tech-

niques and deep learning approaches have been ana-

lyzed on a sporting dataset that contains 109 different

identities in a 30 hours race under variable lighting

conditions, scenarios, and accessories. Our results re-

veal the challenge of performing a runner ReID pro-

cess in these difﬁcult conditions, especially when this

process involves images captured under poor illumi-

nation. Our results also show that deep learning ap-

proaches provide better results than traditional meth-

ods, as they consider the overall context of the image

instead of focusing on individual pixels. Neverthe-

less, there is still signiﬁcant room for improvement

for ReID in this scenario. This work is an interesting

ﬁrst step in this direction, combining different image

processing techniques to boost ReID under challeng-

ing conditions.

The rest of this paper is organized as follows. Sec-

tion 2 discusses the related work in the literature.

Section 3 presents the proposed pipeline. Section 4

describes the considered collection, the experimental

protocol, and the ReID evaluation experiments on the

enhanced images. Finally, Section 5 presents our con-

cluding remarks.

2 RELATED WORK

Many methods have been proposed to enhance images

captured in low-light conditions. In general, illumina-

tion enhancement models based on the Retinex The-

ory (Land, 1977) decompose images into two compo-

nents: reﬂectance and illumination. Some techniques

have been proposed for enhancing images working on

both components (Jobson et al., 1997; Wang et al.,

2013; Fu et al., 2016; Ying et al., 2017; Li et al.,

2018), but they involve a high computational cost.

Guo et al. (Guo et al., 2017) proposed a low-

light image enhancement method that reduces the so-

lution space by only estimating the illumination com-

ponent. First, an illumination map is generated based

on the maximum value of each RGB color channel

for each image pixel. The map is then reﬁned accord-

ing to the illumination structure using an augmented

lagrangian algorithm. Liu et al. (Liu et al., 2022)

proposed an efﬁcient algorithm based on a member-

ship function and gamma correction. This method

also has lower computational cost than other tradi-

tional approaches, resulting in images that do not ex-

hibit over-enhancement or under-enhancement. Since

these methods are explicitly designed for low-light

imagery, they are tuned to enhance underexposed im-

ages. Zhang et al. (Zhang et al., 2019) proposed per-

forming a dual illumination map estimation for both

the original and the inverted images, generating two

different maps. Underexposed and overexposed re-

gions of the image can be corrected through these

maps.

Beyond methods based on the Retinex Theory,

the rise of deep learning has led to the development

of image enhancement methods based on neural net-

works (Cai et al., 2018; Park et al., 2018; Wang et al.,

2019; Kim, 2019; He et al., 2020). Recently, Hao

et al. (Hao et al., 2022) proposed a decoupled two-

stage neural network model which provides compa-

rable or better results than other state-of-the-art ap-

proaches. The ﬁrst neural network learns the scene

structure and the illumination distribution to generate

an image looking close to optimal lighting conditions.

The second neural network further enhances the im-

age, suppressing noise and color distortion.

Low illumination scenarios have been previously

addressed in different computer vision problems. Liu

et al. (Liu et al., 2021) presented an extensive review

of low-light enhancement methods and showed the re-

sults of a face detection task. The authors also intro-

duced the VE-LOL-L dataset collected with low-light

conditions and annotated for face detection. Ma et

al. (Ma et al., 2019) proposed the TMD

L distance

learning method based on Local Linear Models and

triplet loss for video re-identiﬁcation. The method

results are compared with other low-illumination en-

hancement methods in three datasets, two simulated

based on PRID 2011 and iLIDS-VID, and a newly

collected LIPS dataset. Unlike Ma’s approach, which

seeks to obtain person descriptors directly from low-

illumination images, our proposal is more similar to

Liu’s, having a preprocessing stage for the runner im-

ages but in a ReID context instead of face detection.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

642

RP1 RP2

RP3

RP4 RP5

Figure 1: Example of leading runner(s) captured in every recording point (RP1 to RP5).

3 PROPOSED PIPELINE

Our experiments use the TGC2020 dataset (Penate-

Sanchez et al., 2020). This dataset has been used

recently to tackle several computer vision prob-

lems such as facial expression recognition (Santana

et al., 2022), bib number recognition (Hern

andez-

Carrascosa et al., 2020), and action quality assess-

ment (Freire-Obreg

on et al., 2022a; Freire-Obreg

et al., 2022b). It comprises a collection of runner im-

ages captured in a set of recording points (RPs) as

shown in Figure 1. This work focuses on applying

several enhancement techniques to sporting footage

in the wild and analyzing these techniques relying

on the ReID performance. However, to tackle ReID

properly, scenes must be cleaned of distractors: staff,

other runners, the public, and vehicles.

Figure 2 shows the devised ReID pipeline, divided

into three main parts. Firstly, the image is enhanced to

improve the runner’s visibility by applying a low-light

image enhancement method to the input footage, gen-

erating improved footage. As shown in the exper-

iments section, several techniques have been tested

to perform the enhancement process: LIME (Guo

et al., 2017), the dual estimation method (Zhang et al.,

2019), and the decoupled low-light image enhance-

ment method (Hao et al., 2022).

Secondly, we are interested in speciﬁc regions

where runners are located. Once bodies are detected

by applying a body detector based on Faster R-CNN

with Inception-V2 pre-trained with the COCO dataset

(Huang et al., 2017), we have used DeepSORT to

Evaluating the Impact of Low-Light Image Enhancement Methods on Runner Re-Identiﬁcation in the Wild

643

Footage

Region of interest

Low-light image

enhancement

Body cropping AlignedReID++

Runners features

0.141 0.981 ...

0.375 0.864 ...

0.627 0.243 ...

0.345 0.863 ...

0.231 0.381 ...

0.355 0.734 ...

0.687 0.543 ...

0.324 0.263 ...

Enhanced footage

Figure 2: Proposed ReID pipeline. The devised process comprises three main parts: the footage enhancement, the region of

interest cropping, and the feature extraction.

track runners in the scene (Wojke et al., 2017). This

tracker is based on the SORT tracker and introduces a

set of deep descriptors to integrate the appearance in-

formation along with the position information given

by the Kalman ﬁlter used in SORT.

Finally, the embeddings of each runner are ex-

tracted from the resulting cropped bodies using the

AlignedReID++ deep learning model (Luo et al.,

2019) trained on the Market1501 dataset (Zheng et al.,

2015). The resulting embeddings are adopted as fea-

tures to compute ReID results. For every RP but the

ﬁrst one, we probe each runner against the previous

RPs, which act as galleries, thus preserving the tem-

poral progression of the runners throughout the race.

ReID performance is measured using the mean av-

erage precision score (mAP). This metric is ideal for

datasets where the same subject may appear multi-

ple times in the gallery because it considers all occur-

rences of each particular runner. The mAP score can

be deﬁned as:

mAP =

∑

i=1

(1)

Where AP

stands for the area under the precision-

recall curve of probe i and k is the total number of

runners in the probe. In this regard, we have tried

Cosine and Euclidean distances to compute the mAP

values and obtained similar results. Hence we present

results just for the latter.

4 EXPERIMENTS

This section describes the experimental evaluation

adopted and summarize the achieved results. Firstly,

we evaluate runner ReID in the original dataset, i.e.,

without preprocessing, using the proposed pipeline to

set our comparison baseline. We then apply differ-

ent low-light image enhancement methods to explore

their impact on the ReID results. The objective is to

check to what extent the proposed preprocessing vari-

Table 1: Location and recording starting time for each RP.

The reader may observe that the ﬁrst two RPs were captured

during the night. Therefore, just artiﬁcial illumination was

available.

location km starting time

RP1 Arucas 16.5 00:06

RP2 Teror 27.9 01:08

RP3 Presa de Hornos 84.2 07:50

RP4 Ayagaures 110.5 10:20

RP5 Parquesur 124.5 11:20

ants allow to improve the quality of the ﬁnal ReID

performance.

4.1 Original Dataset

As described by the authors (Penate-Sanchez et al.,

2020), the dataset used in this work was collected by

recording participants during the 2020 edition of the

TransGranCanaria (TGC) ultra-trail race, particularly

those taking part in the TGC Classic, where the chal-

lenge was to cover 128 kilometers by foot in less than

30 hours.

Runners start at 11 pm, covering the initial part of

the track with nightlight, roughly eight hours. Win-

ners take approximately 13 hours to reach the ﬁnish

line. The dataset comprises the recording in ﬁve RPs,

two during the ﬁrst eight hours, i.e., with nightlight

conditions, and three more after 7 am, i.e., recorded

with daylight. Given the recording environment, the

images captured at each RP vary signiﬁcantly in light-

ing conditions; see Figure 1. Table 1 shows the dis-

tance from the departure line where each point is lo-

cated in the race track and the starting time for image

recording at each RP.

Although runners wear a bib number and carry

a tracking device, these monitoring systems do not

identify the person wearing them, leaving the door

open to potential cheating (e.g., several runners can

share a race bib to boost the ranking of the bib owner),

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

644

Table 2: ReID results (mAP) for the original dataset, i.e. without applying any preprocessing to the images.

Gallery

RP1 RP2 RP3 RP4

RP2 14.80

RP3 9.11 7.17

RP4 22.66 13.13 19.21

RP5

22.41 20.32 15.09 48.29

Table 3: ReID mAP percentage variation for the enhanced dataset.

Gallery LIME Dual Net-I Net-II

RP2 RP1 +6.75 +5.15 +10.15 +12.18

RP3

RP1 +4.79 +0.61 +9.91 +4.68

RP2 +3.61 +1.59 +11.13 +5.22

RP4

RP1 -0.22 -0.89 +3.25 +1.04

RP2 +3.84 +3.42 +11.20 +10.63

RP3 +0.94 -0.34 +5.65 +3.88

RP5

RP1 +1.86 +0.42 +4.19 +2.85

RP2 +6.02 +4.19 +8.68 +7.15

RP3 +1.69 -0.72 +4.00 +2.83

RP4 -0.28 -3.53 +8.14 +3.02

hence the need to apply ReID techniques in these

long-distance competitions.

For this study, we used only the images of the 109

runners identiﬁed in every RP. As mentioned above,

RP1 and RP2 samples were captured during the night-

time, while the images from RP4 and RP5 were cap-

tured during the daytime.

The images from RP3 provide an in-between sce-

nario, as they started to be recorded in the early morn-

ing hours in a place where the sunlight was blocked

by trees and mountains, creating a situation closer to

twilight than to full sunlight. Once more, the reader

may observe Figure 1, which shows images of runners

in all ﬁve RPs to illustrate the meaningful variation in

lighting conditions between different locations.

Table 2 shows the ReID mAP score using RP2,

RP3, RP4, and RP5 as probe and separately each pre-

ceding RP as gallery. The highest value is obtained

by probing RP5 against RP4 as gallery, since both

RPs were shot in full daylight. On the other hand,

the worst results are obtained by RP3, shot in some

peculiar intermediate lighting conditions and having

only night images available to use as galleries, reveal-

ing the great difﬁculty of runner ReID in the wild.

4.2 Enhanced Dataset

The preprocessed dataset (henceforth enhanced) ap-

plies a previous step to correct the original images in

terms of illumination. In the experiments below, we

do not distinguish between night and daylight RPs to

use the preprocessing. Indeed, it is applied homoge-

neously to any image in the original dataset.

Table 3 provides results for the low-light image

enhancement method (LIME) proposed by Guo et

al. (Guo et al., 2017) and the dual estimation method

(Dual) proposed by Zhang et al. (Zhang et al., 2019).

We evaluated various gamma values in both cases and

found the best ReID results using a value of 0.3. Al-

though it is common in the literature to use more sig-

niﬁcant values, this is usually done with high-quality

images. As a consequence of the wild scenario, the

quality of the images is lower, and the runners are af-

fected by motion blur. Raising the gamma value too

high makes the images noisier and adversely affects

the ReID process.

The LIME algorithm application leads to an evi-

dent improvement in the ﬁrst three RPs, the most af-

fected by the low light. The improvement is less no-

ticeable in the better illuminated RPs, even a slight

degradation in some cases. The results of the appli-

cation of the Dual algorithm are, in general, poorer

than those provided by the LIME algorithm. Both al-

gorithms share the same mathematical basis, but the

LIME algorithm focuses on improving the underex-

posed parts of the image.

In contrast, the Dual algorithm aims to improve

both the underexposed and overexposed parts of the

image. These results suggest that overexposure is not

a signiﬁcant problem for runner ReID when using this

dataset. Consequently, in this scenario it is preferable

to use the algorithm that enhances just the underex-

posed areas.

Evaluating the Impact of Low-Light Image Enhancement Methods on Runner Re-Identiﬁcation in the Wild

645

Original LIME Dual Net-I Net-II

Figure 3: Images captured in RP2 that have been treated with different methods of illumination enhancement.

Opposed to these methods based on the Retinex

Theory, the decoupled low-light image enhancement

method proposed by Hao et al. (Hao et al., 2022) is

a deep learning approach. This neural network does

not apply a series of preset operations according to

an algorithm but has learned to interpret the general

context of the image, leading to better ReID results.

Table 3 shows separately the results of applying

only the ﬁrst stage (Net-I) and applying both stages

(Net-II) of this decoupled neural network. In most

cases, the best results are obtained using only the ﬁrst

stage because it is the stage trained to improve light-

ing conditions. The color and noise corrections ap-

plied by the second stage only provide an improve-

ment in one case, with worse ReID results in the rest.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

646

5 CONCLUSIONS

Runner ReID in the wild is highly complex due to

the long-term component, which does not necessar-

ily preserve clothes unchanged, and the poor light-

ing conditions often encountered. This latter problem

cannot be easily solved by mechanical means, as in-

creasing the exposure time to capture more light leads

to degraded image quality due to motion blur.

To address these difﬁculties, we look for syner-

gies between different methods proposed individually

and applied in origin to different scenarios. Our re-

sults show that applying software methods for illumi-

nation enhancement is a promising approach to im-

prove ReID results. As far as we know, the reported

results are the best achieved without using temporal

information (Medina et al., 2022).

It is a well-known fact that computers do not see as

people do and that the enhanced image that may seem

the clearest to us may not necessarily be the most use-

ful for an automatic ReID process. Figure 3 shows the

nighttime image of two runners with different tech-

niques applied, clearly illustrating that it is not trivial

to identify the best one based on human observation.

Overall, it is pretty likely that no single approach

would be considered the best in all circumstances,

which involves that it would be necessary to choose

the method to apply according to the image features.

Determining the best strategy to achieve this goal

is an interesting topic for future research. In this

sense, it would be of great value for the community

to establish a systematic process for choosing the best

method to apply to a low-light image based on the dif-

ferent quality assessment metrics (Zhai et al., 2021)

proposed in the literature.

ACKNOWLEDGEMENTS

This work is partially funded by the ULPGC

under project ULPGC2018-08, by the Span-

ish Ministry of Economy and Competitiveness

(MINECO) under project RTI2018-093337-B-I00,

by the Spanish Ministry of Science and Inno-

vation under projects PID2019-107228RB-I00

and PID2021-122402OB-C22, and by the

ACIISI-Gobierno de Canarias and European

FEDER funds under projects ProID2020010024,

ProID2021010012, ULPGC Facilities Net, and Grant

EIS 2021 04.

REFERENCES

Cai, J., Gu, S., and Zhang, L. (2018). Learning a deep single

image contrast enhancer from multi-exposure images.

IEEE Transactions on Image Processing, 27(4):2049–

2062.

Cheng, Z., Zhu, X., and Gong, S. (2020). Face re-

identiﬁcation challenge: Are face recognition models

good enough? Pattern Recognition, 107:107422.

Dietlmeier, J., Antony, J., Mcguinness, K., and O’Connor,

N. E. (2020). How important are faces for person re

identiﬁcation? In Proceedings International Confer-

ence on Pattern Recognition, Milan, Italy.

Freire-Obreg

on, D., Lorenzo-Navarro, J., and Castrill

on-

Santana, M. (2022a). Decontextualized i3d convnet

for ultra-distance runners performance analysis at a

glance. In International Conference on Image Anal-

ysis and Processing (ICIAP).

Freire-Obreg

on, D., Lorenzo-Navarro, J., Santana, O. J.,

Hern

andez-Sosa, D., and Castrill

on-Santana, M.

(2022b). Towards cumulative race time regression in

sports: I3d convnet transfer learning in ultra-distance

running events. In International Conference on Pat-

tern Recognition (ICPR).

Fu, X., Zeng, D., Huang, Y., Liao, Y., Ding, X., and Pais-

ley, J. (2016). A fusion-based enhancing method

for weakly illuminated images. Signal Processing,

129(C).

Guo, X., Li, Y., and Ling, H. (2017). Lime: Low-light

image enhancement via illumination map estimation.

IEEE Transactions on Image Processing, 26(2):982–

993.

Hao, S., Han, X., Guo, Y., and Wang, M. (2022). Decoupled

low-light image enhancement. ACM Trans. Multime-

dia Comput. Commun. Appl., 18(4).

He, W., Liu, Y., Feng, J., Zhang, W., Gu, G., and Chen, Q.

(2020). Low-light image enhancement combined with

attention map and u-net network. In 2020 IEEE 3rd

International Conference on Information Systems and

Computer Aided Education (ICISCAE), pages 397–

401.

Hern

andez-Carrascosa, P., Penate-Sanchez, A., Lorenzo-

Navarro, J., Freire-Obreg

on, D., and Castrill

on-

Santana, M. (2020). TGCRBNW: A dataset for run-

ner bib number detection (and recognition) in the wild.

In Proceedings International Conference on Pattern

Recognition, Milan, Italy.

Huang, J., Rathod, V., Sun, C., Zhu, M., Korattikara, A.,

Fathi, A., Fischer, I., Wojna, Z., Song, Y., Guadar-

rama, S., and Murphy, K. (2017). Speed/accuracy

trade-offs for modern convolutional object detectors.

In CVPR 2017, pages 649–656, Hawai, USA. IEEE.

Jobson, D., Rahman, Z., and Woodell, G. (1997). A multi-

scale retinex for bridging the gap between color im-

ages and the human observation of scenes. IEEE

Transactions on Image Processing, 6(7):965–976.

Kim, M. (2019). Improvement of low-light image by con-

volutional neural network. In 2019 IEEE 62nd Inter-

national Midwest Symposium on Circuits and Systems

(MWSCAS), pages 189–192.

Evaluating the Impact of Low-Light Image Enhancement Methods on Runner Re-Identiﬁcation in the Wild

647

Land, E. H. (1977). The retinex theory of color vision. Sci-

entiﬁc American, 237(6):108–129.

Li, C., Guo, C., Han, L., Jiang, J., Cheng, M.-M., Gu, J.,

and Loy, C. C. (2022). Low-light image and video

enhancement using deep learning: A survey. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 44(12):9396–9416.

Li, M., Liu, J., Yang, W., Sun, X., and Guo, Z. (2018).

Structure-revealing low-light image enhancement via

robust retinex model. IEEE Transactions on Image

Processing, 27(6):2828–2841.

Liu, J., Xu, D., Yang, W., Fan, M., and Huang, H. (2021).

Benchmarking low-light image enhancement and be-

yond. International Journal of Computer Vision,

129(4):1153–1184.

Liu, S., Long, W., Li, Y., and Cheng, H. (2022). Low-

light image enhancement based on membership func-

tion and gamma correction. Multimedia Tools and Ap-

plications, 81:22087–22109.

Luo, H., Jiang, W., Zhang, X., Fan, X., Qian, J., and Zhang,

C. (2019). AlignedReID++: Dynamically matching

local information for person re-identiﬁcation. Pattern

Recognition, 94:53 – 61.

Lv, F., Lu, F., Wu, J., and Lim, C. S. (2018). MBLLEN:

Low-light image/video enhancement using cnns. In

BMVC.

Ma, F., Zhu, X., Zhang, X., Yang, L., Zuo, M., and Jing, X.-

Y. (2019). Low illumination person re-identiﬁcation.

Multimedia Tools and Applications, 78(1):337–362.

Medina, M. A., Lorenzo-Navarro, J., Freire-Obreg

on, D.,

Santana, O. J., Hern

andez-Sosa, D., and Santana,

M. C. (2022). Boosting re-identiﬁcation in the ultra-

running scenario. In Proceedings of the 11th Inter-

national Conference on Pattern Recognition Applica-

tions and Methods (ICPRAM), pages 461–469.

Ning, X., Gong, K., Li, W., Zhang, L., Bai, X., and Tian, S.

(2021). Feature reﬁnement and ﬁlter network for per-

son re-identiﬁcation. IEEE Transactions on Circuits

and Systems for Video Technology, 31(9):3391–3402.

Park, S., Yu, S., Kim, M., Park, K., and Paik, J. (2018).

Dual autoencoder network for retinex-based low-light

image enhancement. IEEE Access, 6:22084–22093.

Penate-Sanchez, A., Freire-Obreg

on, D., Lorenzo-Meli

an,

A., Lorenzo-Navarro, J., and Castrill

on-Santana, M.

(2020). TGC20ReId: A dataset for sport event re-

identiﬁcation in the wild. Pattern Recognition Letters,

138:355–361.

Rahman, Z., Pu, Y.-F., Aamir, M., and Wali, S. (2021).

Structure revealing of low-light images using wavelet

transform based on fractional-order denoising and

multiscale decomposition. The Visual Computer,

37(5):865–880.

Santana, O. J., Freire-Obreg

on, D., Hern

andez-Sosa, D.,

Lorenzo-Navarro, J., S

anchez-Nielsen, E., and Cas-

trill

on-Santana, M. (2022). Facial expression analysis

in a wild sporting environment. In Multimedia Tools

and Applications.

Wang, R., Zhang, Q., Fu, C.-W., Shen, X., Zheng, W.-

S., and Jia, J. (2019). Underexposed photo enhance-

ment using deep illumination estimation. In 2019

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition (CVPR), pages 6842–6850.

Wang, S., Zheng, J., Hu, H.-M., and Li, B. (2013). Nat-

uralness preserved enhancement algorithm for non-

uniform illumination images. IEEE Transactions on

Image Processing, 22(9):3538–3548.

Wojke, N., Bewley, A., and Paulus, D. (2017). Simple on-

line and realtime tracking with a deep association met-

ric. In 2017 IEEE International Conference on Image

Processing (ICIP), pages 3645–3649. IEEE.

Ying, Z., Li, G., Ren, Y., Wang, R., and Wang, W. (2017).

A new low-light image enhancement algorithm using

camera response model. In 2017 IEEE International

Conference on Computer Vision Workshops (ICCVW),

pages 3015–3022.

Zhai, G., Sun, W., Min, X., and Zhou, J. (2021). Percep-

tual quality assessment of low-light image enhance-

ment. ACM Trans. Multimedia Comput. Commun.

Appl., 17(4).

Zhang, Q., Nie, Y., and Zheng, W.-S. (2019). Dual illumina-

tion estimation for robust exposure correction. Com-

puter Graphics Forum, 38(7):243–252.

Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., and Tian,

Q. (2015). Scalable person re-identiﬁcation: A bench-

mark. In Proceedings of the International Conference

on Computer Vision.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

648